Machine View Lab
Make the invisible representation inspectable.
Eight local tools demonstrate the human–machine view gap without sending data anywhere. Paste your own benign text or use the instrumented samples.
Safe-lab boundary: examples use ordinary words and the inert marker MT_SAFE_ACK_7F3A. The tools inspect and visualize representation; they do not generate exploit payloads, execute code from samples, or contact external systems.
Unicode and byte inspector
Enumerate code points, UTF-8 bytes, broad Unicode categories, invisibility signals, and normalization differences.
| Offset | Glyph / marker | Code point | UTF-8 | Class | Known name |
|---|
Illustrative confusable lens
Map a small educational subset of common Greek and Cyrillic look-alikes to a Latin skeleton. This is not a replacement for full UTS #39 analysis.
A skeleton match is a risk signal, not proof of deception. Full implementations need standardized confusable data, script resolution, locale context, and tighter rules for identifiers than for ordinary multilingual prose.
DOM text versus rendered text
Compare textContent with innerText on a safe, instrumented fragment containing several legitimate hiding mechanisms.
Benign fixture report
The visible section says the inspection completed normally.
[display:none] MT_SAFE_ACK_7F3A [opacity:0] transparent supporting label [off-screen] archival note Screen-reader-only navigation hinttextContent · structural node text
innerText · rendered text
HTML structural-field extractor
Parse a string without inserting it into the live page. Inventory comments, alt text, ARIA labels, titles, hidden inputs, and data-* values.
Toy linguistic feature decoder
Expose word-length parity and sentence initials. These are transparent educational features, not evidence of a hidden message without a pre-agreed rule and matched baseline.
Detection needs a null model. Ordinary prose naturally contains patterns, short acrostics, parity runs, and stylistic variation. Search over many possible rules creates false positives unless multiple-hypothesis testing is controlled.
Reversible encoding decoder
Decode common machine-readable representations locally. The tool displays text only; it never executes decoded content or treats it as an instruction.
The included samples decode only to MACHINE_TRADECRAFT_DEMO_OK. For attachment-based exercises and copyable AI analysis prompts, use the prompt attachment lab.
Receiver-knowledge simulator
See why “a machine can decode it” is incomplete without the tokenizer, key, model distribution, codebook, parser, or learned convention.
Transformation survival lab
Apply common benign transformations and measure whether a fixed demonstration carrier remains recoverable. Robustness is evaluated against the selected decoder rule, not guessed from visual similarity.
Transformed artifact
Visible diagnostic view
Interpretation: “recoverable” means the selected toy decoder still returns the expected benign value. It does not establish broad stealth, security, or compatibility with an arbitrary model.
Multi-view defense coverage
Select controls present in a document-ingestion pipeline. The score is an educational coverage indicator, not a security certification.
0%No controls selected
Take the demo into an AI chat
Download benign prompt attachments and copy decoder prompts.
The attachment lab packages the same concepts as text, Markdown, HTML, JSON, XML, and a ZIP bundle so you can test a specific AI ingestion path safely.